Abstract
This paper describes a new approach to sub-pixel edge localization by fitting a 3rd-order polynomial surface around each edge point using third order Taylor series expansion and correction with curvature.Sub-pixel edge location is then found as the zero-crossing of the second derivative along the gradient direction across the edge.The zero-crossing of the second derivative along the gradient direction has the same location as the peak of the first derivative.We further propose a new formula to correct the bias to curved edges based on the curvature.The experimental results show that sub-pixel location is unbiased against image noise in the case of straight edges and proposed method has high precision in both simulation images and real images.In the case of straight edges,accuracy of less than 1 twentieth of a pixel can be achieved.
